Shuttle.
On a loom, the shuttle is the small wooden carrier that pulls the weft thread through the warp. In Weft, the shuttle is your AI — the agent you tell things to, which then moves tasks on your behalf.
Weft is the AI task manager for solo founders and small teams.
What the shuttle is, concretely
Any MCP 2025-11-25 compatible client you have connected to Weft: Claude Desktop, Claude Code, Cursor, ChatGPT, Codex CLI, Gemini CLI, Lovable, Replit, Windsurf, Cline, Zed. You can run several at once — each is an independent shuttle moving across the same board.
Under the hood: every client holds its own OAuth token, and the board tags each action with which client performed it. That is how AI-native teams can run Claude, Cursor, and ChatGPT in parallel without the board turning into mush.
What the shuttle can do
Thirty-two MCP tools cover the full board surface — see the tool reference for the complete list. Grouped roughly:
- Read the board —
get_board_state,list_columns,list_tasks,list_sprints. - Find things —
search,fetch. - Edit tasks —
create_task,update_task,move_task,trash_task,restore_task. - Run sprints —
create_sprint,start_sprint,complete_sprint,add_task_to_sprint,remove_task_from_sprint.
Which actions your AI should confirm
Every tool is annotated — readOnlyHint, destructiveHint, idempotentHint — so a well-behaved client knows when to ask first. The tools marked destructive today are trash_task, archive_done_tasks, complete_sprint. Of those, only trashing is reversible, for 30 days via restore_task; a good client still confirms when your wording is ambiguous.
The Model Context Protocol specification defines these annotations as behavioral hints that help clients identify read-only, destructive, and idempotent tool operations.